基于同构查询校准的知识图谱复杂逻辑推理  

Complex Logical Reasoning in Knowledge Graph with Homogeneous Query Revisor

作　　者：汤宇轩 申彦明[1] 齐恒[1] 尹宝才 TANG Yuxuan;SHEN Yanming;QI Heng;YIN Baocai(Dalian University of Technology,School of Computer Science and Technology,Dalian,Liaoning 116024,China)

机构地区：[1]大连理工大学计算机科学与技术学院,辽宁大连116024

出　　处：《中文信息学报》2024年第9期24-35,共12页Journal of Chinese Information Processing

基　　金：国家自然科学基金(U1811463,62072069)。

摘　　要：知识图谱复杂逻辑推理是知识图谱中的一项重要任务,其目的是根据给定的起始节点和逻辑表达式来推理出答案节点。先前的工作主要关注的是如何对实体、关系和查询进行建模,忽略了相似查询对当前查询的影响。因此,该文提出了一种相似查询的定义(称之为同构查询),并设计了一种基于同构查询的组件,它可以利用同构查询的特性,在推理的每一步缩短查询嵌入和答案嵌入之间的距离,在不改变原有复杂逻辑推理模型结构的基础上提升模型的性能。实验结果表明,该文提出的组件可以在不同的数据集上为各类不同的基线模型带来1.6%-3.3%的提升,证明了该方法的有效性与灵活性。Complex logical reasoning is a fundamental task in knowledge graph,which is to deduce the target entities from anchor entities and relations in query.However,previous approaches haven’t considered relevance among similar queries that can be helpful for inference.This paper designs homogeneous query revisor,a flexible component which can be added to current approaches without changing their structure and improve the performance of complex query answering based on the features of similar queries.The revisor can bridge the distance between query embedding and answer embedding in each reasoning step.The performance of this revisor on different data sets is 1.6%-3.3%higher than that of the baseline models,which demonstrates the flexibility and effectiveness of our approach.

关 键 词：知识图谱 复杂逻辑推理 同构查询 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]